This content provides an in-depth, long-term perspective on using OpenClaw, an always-on AI agent, moving beyond initial impressions to showcase practical, daily use cases, evolving workflows, and the realities of its limitations after over 50 days of continuous use.
Mind Map
Clique para expandir
Clique para explorar o mapa mental interativo completo
Most OpenClaw videos right now are either first-week impressions,
or set up tutorials, or showing use cases after three days of usage.
Nobody can tell you what happens after the first month, because
they haven't been there yet.
I have.
Every single day. For over 50 days. Through every single
iteration of this tool:
when it was ClawdBot, when it was MoltBot (which I refused
to call it that) and when it is now OpenClaw. I made the
setup video that ended up in the official OpenClaw documentation.
I built Clawdiverse, the community directory of use cases, and the
most common post on Reddit is still:
"I set up OpenClaw, but I don't know what to use it for."
This video is the answer: 20 real use cases from my daily life, plus the
honest truth about what breaks, how it breaks, and how to deal with it.
Quick context for anyone new.
OpenClaw is an always-on AI agent that runs on your server, your VPS, a
Mac Mini or even a Raspberry Pi, 24/7.
It connects to your messaging apps that you already have
on your phone, like Telegram, WhatsApp, Discord, iMessage.
If you need the full setup, I have a video for that
(link in the description)
This video is about what you *do* with it once it's running.
And a super important thing:
Every single prompt for every use case I'm about to show
you is in this document
(also link in the description)
It's ready for copy-pasting and using with your own agent.
Let's go see how it all works.
Before the use cases, let me walk you through what
50 days actually look like.
Because the way you use this thing in week one is
nothing like in week seven.
Week one is novelty.
You're asking it random questions, testing what it can do, kind
of using it like a ChatGPT.
But one decision I made from day one saved me over and over in
coming weeks markdown-first.
A lot of people build their workflows around SQLite, databases,
vector stores, custom schemas.
I put everything in Obsidian from the start in plain text files.
Any person can read them, I can read them.
Any program can work with them.
It's just plain text when the next thing after OpenClaw or the next
iteration of OpenClaw comes along.
My data moves with me in five seconds.
No lock-in, just files.
I can do anything with them.
By week three, you start building automations, warning
briefings, background checks.
It starts being more useful and you can feel it in week five when
you start using it more and more.
You hit a wall, everything is in one conversation and
everything is mixed together.
Research, bookmarks, analytics, daily tasks, and there's more and
more context pollution, and that's when I learned to separate contexts.
I now have one Discord channel per workflow.
This way, research doesn't bleed into analytics.
Bookmarks don't pollute daily assistant tasks, and I'll show you the
full architecture later in week seven.
Another lesson.
Not every channel needs the same brain.
You need to match the model to the task.
So Opus is for deep thinking and cheap models are good enough for
routine work, and that's when costs stop being scary and crazy.
And by week eight and onwards, it stops being a chatbot and becomes
a system and lesson from this.
Is three important principles after 50 days that I would recommend everyone
is to have everything in markdown from the beginning to separate context
and to match the model to the task.
Here's what we are covering, 20 use cases across six categories,
and I'm going to move fast, real screenshots, real conversations,
real results, and if you only steal three ideas from this entire video.
I'll tell you exactly which three closer to the end of
the video, so stick around.
It's gonna be worth it.
Most of my setup runs in Discord now, which wasn't the case from
the beginning, and I'll show the channel architecture and model
routing later for now, starting with the things that run every single
day without me touching anything.
Every morning at 7:00 AM my agent scans a bunch of tweets
from accounts they follow.
Picks the top 10, writes them to my Obsidian notes.
Appends any video ideas to my shipping backlog and sends me a summary.
I wake up and I don't need to scroll through the feed
to know what happened.
The most important and interesting part is already waiting for me
and it's tailored to my interests or what I'm currently working on.
So today's stories are about Anthropic and OpenClaw,
new Gemini model dropping.
And this is already saved in Obsidian, and a couple video
ideas are also added to a note in Obsidian, and the value compounds
with time because it doesn't just summarize, it connects the dots.
Like, Hey, this tweet about model pricing connects to
your video idea about cost optimization, that kind of thing.
But briefings are the gateway drug.
Everyone starts there.
Let me show you what comes after.
This is my favorite use case.
Every morning my agent fetches Wikipedia's 'On This Day' events,
picks the most impactful historical event and then generates a woodcut
style image showing 10 seconds before that event happened, like the
iceberg approaching the Titanic or the apple, about to fall on Newton's
head, this kind of stuff, and then it pushes to my TRMNL e-ink display.
In a mystery mode, it only shows the date and location,
and I need to guess the event.
Sometimes it's obvious, sometimes it's a lesson for me.
So for example, this is February 1st, 2003 over Texas.
This is seconds before shuttle Columbia Disaster.
Where it blew up, and these are the kind of images that it produces.
For example, this is just before the last public appearance of
the Beatles on the rooftop.
And this is the beheading of Mary Queen of Scots, which changed
the course of Europe's history.
And this is part of my daily ritual.
Now I walk past the display, look at the new picture, try to guess what it
is, learn something new about history.
Every single day a new one is waiting for me.
This is maybe less of a use case and more like part of my daily routine and
also an important part of the routine.
So I have two cron jobs that I now never think about every day at 4:00
AM when I'm sleeping, my agent updates its own skills from ClawHub updates.
The OpenClaw package itself restarts the gateway, and
then reports the results.
When something breaks during an update, it tells me and every day,
half an hour later, a separate cron job backs up everything
important, all configuration files, workflow directory, crown
schedules, SOUL file, MEMORY files, skills, everything that defines how
my agent works or who he even is.
So if my server dies tomorrow, I'm back up in half an hour, not
rebuilding from scratch, not trying to remember how I configured things.
Just restore and go, and that's the whole point.
It just runs.
And if something goes wrong, I can recover easily.
And just as a reminder, all of the prompts to achieve all these
tasks and use cases are in this.
Just on GitHub, so you can just go there, copy paste to
your agent, and be good to go.
Always on checks, like for example, background health checks.
This used to feel like the headline feature, but now I think of it as
background guardrails useful, but it's only one piece of the system.
My agent runs routine heartbeat checks every 30 minutes.
And I can define what it does every 30 minutes.
For example, it scans my emails, checks my calendar, monitors
my services, and it catches things I would have missed.
For example, one time it just sent me this message out of the blue
about a Netflix payment failure, and I had no idea about it.
It found it during a routine email scan.
I paid it five minutes later so I could safely watch more TV shows.
Or rather, I don't really watch TV shows.
I don't have time.
Rather, my kids could safely watch more hunters and catches a lot of
other things like domain renewal coming up, or a meeting I was about
to miss, or a relevant newsletter article that it found during a Sunday
heartbeat scan that connected to a video that I was currently working on.
And none of these were tasks that I assigned.
My agent just found them.
It knew what I'm working on and the things that would normally
fall through the cracks.
That's exactly what gets caught by the agent.
And an important piece of setup here is that the agent is in
a draft only mode for email.
It can read my inbox.
It can flag what's important.
It can even draft responses, but it cannot send it.
I need to review and send the email.
And, and it doesn't happen without my approval.
For now, there's no robust general solution yet for
prompt injection via email.
So I treat inbox content as potentially hostile research
and content creation.
And I love the research part of working with OpenClaw.
It just appeals to me to do it on the go when I have ideas
and I want to discuss it.
For example, for this video, I told my agents to research what people
are doing with OpenClaw, and then it spawned five parallel sub-agents.
One Search Twitter, one crawled Reddit, one hit Hacker News.
One analyzed YouTube competition and one scraped multiple different forums.
They all ran simultaneously and produced massive
structured research files.
Each one of them with competitive analysis ranked video ideas, full
outlines, with source links, and it took them minutes, not hours.
The research files for this video alone produced by these
agents are over 50 pages.
And it gave me a clear understanding of what people are doing.
And more importantly, not yet doing with OpenClaw and what I need to
focus on to give the most value.
And it's not just for video research, it's for any kind of research.
You can define how broad it should go.
How fresh the data needs to be.
Is it the last seven days?
Is it the last three months?
And then let it run and come back with results so you don't
start from scratch and you have a solid base to start with.
And speaking about YouTube, I have two dedicated Discord channels
related to YouTube creation.
The first is my YouTube analytics channel.
It has access to all my stats via the API, and I can query anything in.
Natural language, like how did my last five videos compare on retention or
which topics get the most engagement?
Or compare my OpenClaw videos to my Claude Code videos, and
then it slices and dices the data anyway I want on demand.
It's like much more flexible than YouTube studios built in dashboards,
and it also synthesizes the numbers.
And gives ideas and advice based on that.
You're not getting this from any kind of analytics.
For example, this was just a random question how my last two weeks went,
and then it gave me the view numbers.
It showed me when I published the videos to understand where was the
spike, how it goes down, where is the uptick, what were the key insights?
What was the watch time, which I didn't even ask about, and
what was the bottom line?
So everything related to what I care about, which I didn't even ask.
It just knows and gives me the details.
As granular as I want them to be, and the second channel is
my video idea research channel.
Throughout the week, I drop links, articles, tweets, half warm thoughts,
what I want to include or what I don't want to include in the next video.
And then the agent goes and enriches those snippets.
It connects the dots across different sources and builds context over time.
And by the time I sit down to work on the video, I don't start from scratch.
I have weeks of accumulated organized research waiting for
me, and this exact video is a great meta example of it.
The research for this video accumulated over probably like
three weeks, maybe even more.
Links from Reddit, insights from Discord, competitive
analysis, audience pay points, all fed over time.
As ideas came to me as I got some data, as I got some numbers.
And then everything got organized and connected by the agent so that I
have a solid baseline to start with.
And having those two different channels for YouTube are important
because the separation matters, analytics context, stays isolated,
and then research builds over weeks without polluting other conversations.
Next is summarizing practically anything.
Throw any URL at it, like an article, a YouTube video, a research paper,
A PDF, and you get a summary back and I use it multiple times a day.
For example, let's summarize my last video.
/summarize and a URL and it goes to work, and a few seconds
later I get the summary.
What's the video about?
What's the core message?
What are the key numbers?
What are the practical steps?
And from there, I can tell it to either expand or this is good enough,
or maybe save it to an Obsidian note.
I can do anything.
Infrastructure and DevOps.
My agent migrated me from the alt ClawdBot package to OpenClaw.
It found both packages running at the same time.
It killed a zombie process running at 160% of CPU.
It deleted the old system services fixed seven days of silently broken
cron jobs that went unnoticed and all from one message, go fix everything.
It's also connected to my VPS server via API, and I host everything there.
And the first time I connected it, it reviewed more than 20 apps.
Running there, flagged some unhealthy services.
Did some fixes and right now I can do anything with it.
I don't need to SSH to my server.
I can check the health of the whole server or specific apps and it can
fix the apps, restart the apps.
I have basically like a remote control to my whole server just in my
Discord, and again, about security.
There is an allow list of commands that it can do, and there's
also a deny list of commands that it cannot do by itself.
It has to ask my permission first.
Works pretty well and I haven't had any issues so far.
It allows me also to code from my phone.
I can tell my agents to go fix a bug.
Build a feature, create a PR, all from my phone while I'm away from my
desk, and you don't need your laptop because your AI has your laptop.
But to be completely honest, I do not use it for production
as my main way of programming.
I only use it for some quick fixes or simple ideas that come to my mind,
and I wanted to test them on the go.
For my main workflow, I still use Claude Code and Codex on my
desktop Daily Life assistant.
And to start with, what everybody's doing is email, triage and draft
replies beyond the proactive catches that I already showed you.
The day-to-day email workflow is simple that it reads my inbox.
Flags what's important and drafts responses.
I review and send.
This way I can easily reply to an email the same day rather than
during the weekend when I have more time to go through my inbox.
This one is useful.
Calendar and family management, not just for
myself but for family as well.
For example, I have a group chat on WhatsApp this time.
Because my wife doesn't use Telegram or Discord, and in
WhatsApp group, I can drop messages like Schedule Dentist
Thursday at 3:00 PM and it's done.
And my wife can add events, check the schedule, get
reminders all through that same group chat and chat interface.
It is simple, but it is helpful as well.
And once it works, you start asking like, what else can you do?
Voice note transcription.
This is something that I always thought I would be using on my phone
because I can dictate stuff and it's there, but when I tried it, I never
went back to listen to those messages.
Now it can actually be done automatically.
So for example, I send a voice message on WhatsApp, telegram or
Discord, and it transcribes it with Whisper model and responds in
text, quick thoughts while driving.
Shopping lists, while walking, meeting notes on the go.
I just talk and then it handles the rest.
And if it's something important, it puts it in maybe a file in Obsidian.
Or sends a message or drafts an email so that I don't even
need to go back to those notes.
They are already taken care of and more small stuff from daily
life, like find me a good coffee shop within walking distance.
It uses Google Places API, so it has access to ratings reviews.
Walking distances from my home, what people like or dislike in that place.
What are the prices?
So it can do much more in seconds than I would do myself, and it's
helpful for one of searches as well.
I was searching for snowboard boots to go snowboarding, and
I have a large foot, so it's an issue to find my size.
So it serves multiple shops.
Where I could buy one, it checked online whether they had
my size and it gave me three options where I could go and buy.
So I went to the first one, bought one, which were not available anywhere
else in the city or almost, and.
Took me 20 minutes for weather forecast.
It doesn't tell me just like tomorrow is gonna be sunny.
I have it on my watch, but it did warn me once, for example,
about minus 19 degrees cold snap coming up, which is like minus
two, minus three Fahrenheit.
And that was quite helpful.
And then there are reminders about maybe rehab exercises
with snooze capability meeting reminders before my weekly calls.
These are all small things on their own, but they do add up and they are.
Genuinely helpful in everyday life, helping friends with their
technical issues and problems.
And this one is personal.
So my friend wanted to set up his own OpenClaw after watching my videos,
and I added him to a WhatsApp group with my own agent, and myself and my
agent spent three and a half hours.
Guiding him step by step through the entire setup in a non-English
authorization, debugging like the whole saga, and it was all done
either via asking questions or posting screenshots in the group chat.
And my friend would take a screenshot of an error.
My agent would read it and explain the fix, tell what the next
step is, and that's how it went.
Well, previously, I would have had to answer everything myself.
Now my agent answered 90% of the questions, and I just added
context from my own experience to some of the answers.
After a few days when my friend installed his own instance, the
questions become more rare, and then they stopped because he
switched to asking his own agent, and I didn't have any technical
questions from him in weeks.
And I only hear updates from time to time as to what kind
of automations he's able to do.
And for a non-technical user who runs an accounting company,
I'm genuinely amazed how quickly and how far he has gone already.
I also have a group chat with my wife and my agent.
And it adds fun to our conversation, some jokes or second opinion on
things that we are discussing.
And once my wife texted in that channel to tell Rad to
play with me when I was too busy talking to my agent and.
I left what I was doing and we played.
And the last part is about my evolution of how I am
using OpenClaw today.
So migrating to Discord, building the knowledge base and some fun projects.
Migrating to Discord is one of the biggest changes in my setup
over the last 50 days, and I think it's the most important
thing that I can share with you.
So first I started on WhatsApp and then quickly moved to
Telegram because it just had more features and was just better.
And most people started on Telegram too.
But around week five, I started hitting a wall.
Everything was in one conversation.
My YouTube stats were mixed with my bookmarks.
My research was mixed with my daily assistant tasks.
And context was getting polluted and it was hard to start finding
things or getting back to something that I saw previously.
So I migrated to Discord and now it's night and day difference.
Instead of one conversation or multiple separate agents like I had
before, I now have channels and each channel is a dedicated workspace.
With its own context, there's a channel for YouTube analytics, a
channel for video idea research, and inbox channel for bookmarks, a general
channel for daily assistant stuff.
And each one stays focused on just that one task and area.
And the important part is that I can set different models per channel.
For example, my YouTube stats channel can use a cheaper model
because it's mostly data retrieval.
My research channel uses Opus because I need deep thinking during research.
My inbox channel uses a fast cheap model because it's just.
Processing links, summarizing and categorizing, and that's
how we keep costs down.
Matching the model to the task, not the other way around, and not
overpaying for expensive models for tasks that can very well be
done by a cheaper, smaller model.
And switching to Discord wasn't about the app itself.
It was about the architecture.
I would prefer staying on Telegram or WhatsApp, but this works better.
It separates contexts.
It has cleaner conversations, it has better formatting, and you
have per channel models, and you have lower costs because of that,
and with this setup of channels, I always know where to go for what.
And I'm careful about adding more channels.
I'll probably have more with time, but I only add a new one as and
when I really, really need it.
And that's what 50 days looks like.
You stop using the tool and start designing how you interact with it and
start getting more value out of it.
Bookmarks, bookmark is a fun.
I used to use Raindrop for bookmarks.
I had a paid subscription.
A separate app, manual tagging, putting everything into folders.
And as I was using that, I even built a system that was
regularly pulling the bookmarks from Raindrop using the API
and putting them in my Obsidian.
What's more, I even built a skill for OpenClaw to manage.
Bookmarks from Raindrop and to enrich them and to create tags
automatically and to pull to Obsidian, just the whole thing.
But at some point I realized that I don't need an
intermediary, which is Raindrop.
I can do it myself directly.
So now I just drop any link into my Discord inbox channel, uh,
and the agent does the rest.
It summarizes the content.
It extracts key information.
It creates text for every bookmark, and over time builds a knowledge
graph, connecting related links, all in markdown, all searchable,
all building context over time, all in Obsidian, and it runs on a
cheaper model because link processing doesn't need Opus and I canceled
Raindrop and I don't miss it.
Obsidian and knowledge base.
As you might have noticed already, most of the things that I do with
my agent in some way or in another end up in Obsidian, and here's where
the markdown first thing pays off.
I have almost 3000 notes in Obsidian, and my agent indexes all of them every
night using QMD for semantic search.
Semantic search means that I can ask questions like,
what did I decide about?
Thumbnail design last month, and then it finds the exact note.
It's not keyword matching, it's semantic understanding
what I'm asking about.
So for example, what's the overarching theme in my last five videos?
It lists the five videos and says that the overarching
theme is making AI coding tools actually useful in practice.
It's not something that can be found by searching for keywords.
And as I work more and more with my agent, as I forward random
things, random thoughts, random links, ideas throughout the day,
they go into Obsidian as markdown files and the agent organizes them,
and then every night at 3:00 AM.
The entire index rebuilds, and then I have this nice
semantic search available to me.
When I first set this up, it took a few minutes to build the
initial embedding index because it was about 3000 or then two and
a half thousand nodes, and now it updates automatically every night,
and it takes about 10 seconds.
People call this second brain stuff.
Mine is always on, does all the organizing for me and everything is in
plain text files that I own forever.
No database, just markdown files and semantic search on top of it.
This one is funny.
Like one day I was checking out my analytics on my website, velvet
shark.com, and I noticed that there were many hits to a WordPress
logging page, but my website is not on WordPress, so I asked my agent
to set up a honeypot on my website, like a fake WordPress logging page.
That re rolls anyone who tries to log in and it built the page, created a
full pull request and deployed it.
And when I went to this page, it's not a WordPress page, it's a fake
page that you can enter anything you want in here and try to log in.
Yeah, you get the idea.
To be clear, this is purely on my own domain, catching bots that
scan for WordPress admin pages and try to hack into WordPress, uh,
websites Don't use this pattern to impersonate any real services.
I'll probably have it for a few days or a couple weeks, and then I'll
get bored of it and I'll remove it.
But it was fun to just think of something like type a few
sentences on my phone and have this fun, just like party trick.
Live on my website, so one minute the agent is managing your
infrastructure and the next, it's setting up some like fun pranks,
and that's the fun part of it.
My agent also can create diagrams and graphs automatically through
the Excalidraw MCP integration, like architecture diagrams,
flow charts, concept maps, it generates them on the fly.
During conversations, and then I only need to change a few
things, maybe move stuff around and maybe add a few words.
Like for example, this chart that was presented earlier, it
was done completely by my open cloud instance for this one.
I didn't have to change anything at all.
Oh, maybe I changed the colors of the models here.
Nothing else.
So if you need more than just creating text or some copy and you
need to visualize a workflow, you just ask and it draws for you home
automation, this one is in progress for me currently and I'm showing
it because it's where my setup is heading next as I'm setting up home
assistance for smart home control.
I have two home assistant voice preview edition
devices for voice control.
So full home automation can be managed through OpenClaw, like
light control climates, routines, all through chat or voice, not even
having to write anything, and it's closer to what Siri should have been.
Than anything that Apple has shipped so far.
So that's 20 use cases from my daily life, but I'm not the only one.
The community is doing like incredible things.
People are running like actual businesses through their agents.
Customer quoting, invoicing, lead generation deal, closing, SEO, and
people are managing smart homes with home assistant controlling 3D printers
connecting their cars even like their Teslas, and they're making phone
calls through their voice agents.
Connecting robots with cameras, fact checking conference speakers
in real time, and even deploying code from the Apple Watch.
I haven't done that yet, but sounds interesting.
And I built Clawdiverse.com.
To catalog all of it, link in the description and the range
is wider than I expected.
And if 20 use cases that you already saw is not enough, you can
find many more on Clawdiverse.com.
But this video is about my own experience.
So let me tell you what nobody else will.
But before going to the honest part.
If you installed OpenClaw today and you are overwhelmed, start with
these three things, one draft only.
Email triage With urgent alerts, it catches things that you miss.
Two, a daily briefing that writes to a markdown file.
Morning context, organized automatically for you
as you start the day.
And three one Discord inbox.
Channel four bookmarks.
Drop links.
Agent enriches them, replaces a paid app immediately and builds
your knowledge base over time, which is more and more valuable as
you go do these three things for a week and you'll start getting it.
Everything else grows from there.
Okay.
What doesn't work well?
No sugar coating.
Number one, memory loss and context.
Compaction, my agent forgets things sometimes mid conversation.
No warning, like it just drifts away for a second.
It's not there for a while.
Then replies to something from three sentences back and
then says it's still there.
Everything is fine.
And this is the number one technical frustration that
people mention everywhere.
Silent compaction.
The context window fills up.
The agent compresses the conversation and important details disappear.
The mitigation is to write everything to files.
Use QMD for semantic search.
Use compact manually before the system does it automatically.
But it's still rough chat.
GPT at least warns you.
When context is getting long OpenClaw just silently
compresses and moves on.
You can at least type a status command to see how much context
is left, but that's not ideal.
For example, here it shows that I have.
35% of context used up and no compactions yet in this session.
If I see that context is filling up and it's more than 50%, I can start
a new session and then it starts with a fresh context from zero.
Next, the cost reality, I covered this in depth in my cost optimization
video linking the description.
The quick summary is that Opus is amazing, but expensive, and
the answer is multi-model routing.
Use opus for the real thinking and cheaper models for
heartbeats or sub agents.
And my Discord channel setup is built around this.
Each channel uses the model that.
Matches its task.
And you can also set different models for heartbeats or other simple tasks.
And my cost calculator, where you choose different models for
different primary model or heartbeat model or sub-agent model shows how
much you can save and it's real money and you need to plan for it.
Another problem is, what do I use it for?
Problem.
This is the most common post on subreddit.
It's what people ask in Discord.
It's what people post on Twitter.
And the harsh truth is that you need to realize one thing.
If you don't have workflows to automate, OpenClaw won't
invent them for you if you don't manage your calendar.
And AI calendar manager won't help if you don't check email ai.
Email triage is pointless.
So.
It can help you if you have what to help with, and the people getting
the most value already had systems OpenClaw made their systems
easier, faster, and automatic.
But the systems were there that said, this video is the
answer to, what do I use it for?
I just showed you 20 ideas plus a starter pack.
Pick three, start there, grow from there.
Next is tasks that need babysitting, like complex, multi-step tasks,
still fail or need pushing.
Browser automation is still flaky.
Sessions, extensions drop.
The agents sometimes go silent mid-task, and you have to
ask like, Hey, how's it going?
It works better as an assistant than an autonomous agent, at least for now.
The simpler the task, the more reliable it is, and the more complex,
the more you need to check in and to give more detailed instructions
for losing context and for compactions, it helps when you
explicitly tell to launch subagent.
This tweet from Matt shows it nicely.
So this is the main agent, the orchestrator, and it is launching
subagent, and each subagent gets its own context window.
So while doing research or performing tasks.
Those sub-agents do not into your main context window, and your
main agent only does coordination instead of all the work.
And this works wonderfully.
And this is also how I do all, almost all my research where the
orchestrator launches subagents and then just like gathers.
All the content, all the research, and then synthesizes the results
and gives the final output.
Next, and I will never stop highlighting this.
Security is real.
There is no full proof general solution for prompting injection yet.
So I treat inbox content as hostile or potentially hostile.
And if your agent reads untrusted emails.
Someone could craft a message that makes your agent do
something that you didn't want to.
The way I solve it is not exposing anything to the outside world in
terms of connectivity and having all my machines on Tailscale.
Then draft only email mode approval needed for destructive actions.
And running security audit regularly for security audit.
Either use this from ClawHub, ClawdBot security check, or just point
your agent to this security page in OpenClaw documentation and.
Tell it to implement and verify everything on this page and if
something is not done on your end to implement these security measures.
And lastly, my own failures.
I want to be specific about things that went wrong for me.
And to be honest, most of these were closer to week one than week seven.
Because the whole system is evolving and improving rapidly.
But for example, daily update.
CR job was using the old ClawdBot commands after the migration to
OpenClaw failed silently for days.
I didn't notice and missed a few updates because of
just a simple package.
rename. Authentication, debugging with my friend.
Three plus hours of false starts, credential comparisons,
complete reinstalls.
The setup is sometimes genuinely hard, but luckily, my own agent
was doing 90% of debugging, so at least that helped.
Context.
Compaction hit me a few times in the middle of a complex research task.
Now I'm more aware of that.
I'm mitigating that, so I'm getting it less and less.
The Discord migration itself took also some iteration, getting the
right channel structure, figuring out which models work best,
where, how to set up different models to different channels.
Migrating context from Telegram conversations.
All of that wasn't like super easy or instant.
It took about a week of tweaking to get everything right and none of
this made me stop using OpenClaw.
But nobody else is telling you about this stuff.
Everything seems rosy, which is not always the case.
Okay, so let's do some scoring.
Let me put a number of it from a few different perspectives.
In terms of setup difficulty, it's about seven out of 10, which is
intentional because it's still a work in progress and it might be dangerous
if someone doesn't know what to do.
So it's intentionally not being done easier.
So if you know how to go through, set a process, you are
at least somewhat technical and you are aware of the dangers.
Daily value once running and once you have your own setup tailored to
your needs, easily nine out of 10.
Reliability for simple workflows.
Eight out of 10.
Very good.
Four complex, especially browser tasks.
Still hit or miss five out of 10, and there is a lot still to improve there.
Best feature, at least for today for me, is Discord channel architecture.
With per channel models, the biggest unlock is file-based memory with
markdown-first files with nightly semantic indexing and retrieval.
And I'm building up my knowledge base day after day, and it's more
and more useful for me with time and it's only getting better.
Most quietly useful is background, heartbeat checks.
You can start simple and as you think what else you might be checking.
Every once in a while you can add more to your heartbeat.
And the biggest pain for now is still memory and context compaction.
What surprised me is first that it gets better over time.
The more context it has in its cell file and in its memory file and
folder, the better it understands you.
And after 50 days it anticipates what I need, but even internalized like
tiny style preferences over time.
The shark emoji, the language switching between dms and groups, it
learns you over time and it can easily surprise you how well it learns you.
So when you start in week one with what seems like a simple tool,
by week three, it feels almost like an infrastructure already.
And by week seven, you reorganize your workflow around it.
And that's when it stopped being a chatbot for me and became a system.
Another surprise for me was that it replaced more than expected.
I expected it to replace ChatGPT, but it also replaced parts of Zapier,
raindrop parts of YouTube, studio analytics, parts of web analytics,
and half of my Apple shortcuts.
And for personal use, I'm not paying for Zapier anymore.
I'm not paying for Raindrop anymore, and I don't miss either of them.
And the ecosystem is not just growing, it's exploding.
There are thousands of skills on ClawHub.
There are hosted services launching for non-technical users, which
still might not be the best idea, but it's there and the tooling and
security is maturing very fast.
And when the security will be much harder and when the setup gets
eventually easier, everyone will be running some version of this and
the capability is already there.
The onboarding isn't by design for now.
So would I recommend OpenClaw?
But with some conditions, yes.
If you have workflows to automate.
And you are comfortable with a terminal and you understand
the cost implications not yet.
If you want something that just works out of the box, you are not technical
or you expect fully autonomous AI that never needs babysitting and
does stuff for you start to finish.
I feel like we are currently using maybe 5% of what this can do.
And the ceiling is like absurdly high, but the floor
still has some holes in it.
And if you are okay with this trade off, if you like
building towards something.
Then this is the most fun like I've had with technology in years, and I,
I keep having fun every day with it.
So that's my journey.
50 plus days every day through ClawdBot to MoltBot, to
OpenClaw, rebrand saga.
Now OpenAI and Foundation Shift.
I've seen the community grow from a few hundred to tens of thousands.
And I've seen my bot fail.
I've seen my bot kill itself.
I've seen it forget what it was doing, but I've also
seen it migrate my server.
Research the entire video with parallel agents, help my friends set
up for three hours, or generate art that makes me smile every morning.
So for me, I'm not going back and that's the strongest
endorsement I can give.
As always, all links are in the description.
There are always all the prompts for all the use cases,
some bonus security tips, cost optimization tips, my setup video.
If you want to get started, the cost calculator, the official
documentation, and once you watch it and start doing and
using those prompts, drop your favorite use cases in the comments.
I want to hear what you're building, what you are trying
or maybe failing, and subscribe.
If you want more practical ways to use AI and now go build yourself
a system and have fun doing it.
Clique em qualquer texto ou marcação de tempo para ir diretamente àquele momento do vídeo
Compartilhar:
A maioria das transcrições fica pronta em menos de 5 segundos
Cópia com um cliqueMais de 125 idiomasPesquisar no conteúdoIr para marcações de tempo
Cole a URL do YouTube
Insira o link de qualquer vídeo do YouTube para obter a transcrição completa
Formulário de extração de transcrição
A maioria das transcrições fica pronta em menos de 5 segundos
Instale nossa extensão para Chrome
Obtenha transcrições na hora sem sair do YouTube. Instale nossa extensão para Chrome e acesse com um clique a transcrição de qualquer vídeo direto na página de reprodução.
